hermes - 💡(How to fix) Fix [Bug]: xAI Grok (xai-oauth) returns HTTP 400 or streaming fallback errors on Responses API [1 pull requests]

Bug Description

Using Grok models (grok-4.3) with provider: xai-oauth and api_mode: codex_responses fails after hermes update.

Two distinct failure modes occur:

1. Streaming path: responses.stream() fails with runtime errors about missing terminal events, followed by a broken create(stream=True) fallback.
2. Non-streaming path: responses.create() returns HTTP 400: Client specified an invalid argument when MCP Firecrawl tools are present.

Minimal direct requests to https://api.x.ai/v1/responses succeed with the same credentials, confirming this is not an OAuth/entitlement problem.

Steps to Reproduce

1. Configure Hermes with:

   model:
     provider: xai-oauth
     default: grok-4.3
     base_url: https://api.x.ai/v1
     api_mode: codex_responses

2. Ensure MCP Firecrawl tools are enabled (default).
3. Run any prompt:

   hermes chat -q "Reply OK only." -Q

Expected Behavior

• Request succeeds using the Responses API.
• Logs show codex_response_request and a normal turn completion.

Actual Behavior

Either:

• Responses create(stream=True) fallback did not emit a terminal response.
• or

  HTTP 400: {'code': 'Client specified an invalid argument', 'error': 'Invalid arguments passed to the model.'}

Affected Component

CLI (interactive chat)

Messaging Platform (if gateway-related)

No response

Debug Report

Report       https://paste.rs/xq8PR
  agent.log    https://paste.rs/qdhwe
  gateway.log  https://paste.rs/wSdK7

Operating System

Arch Linux

Python Version

3.11.15

Hermes Version

v0.14.0 (2026.5.16)

Additional Logs / Traceback (optional)

Root Cause Analysis (optional)

1. Streaming path always chosen for codex_responses

In run_agent.py, _interruptible_api_call() unconditionally routes codex_responses through _run_codex_stream() → active_client.responses.stream(...).

xAI’s Responses streaming implementation is incompatible with the OpenAI SDK’s state machine. The SDK raises RuntimeError (missing response.completed / prelude errors), triggering the broken create(stream=True) fallback that also fails.

2. Tool schemas contain regex pattern constraints that xAI rejects

The MCP Firecrawl tool mcp_firecrawl_firecrawl_search registers schemas with:

"includeDomains": {
  "items": {
    "type": "string",
    "pattern": "^(?=.{1,253}$)(?:[a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?\\.)+[a-z0-9][a-z0-9-]{0,61}[a-z0-9]$"
  }
}

The existing strip_pattern_and_format() sanitizer in tools/schema_sanitizer.py only walked OpenAI-style tool shapes ({"function": {"parameters": ...}}). It never processed Responses-style shapes ({"name": "...", "parameters": ...}).

xAI’s /responses endpoint rejects pattern (and format) keywords in tool schemas, producing the 400.

Proposed Fix (optional)

• Make the streaming decision provider-aware: treat xai and xai-oauth as non-streaming for codex_responses.
• Extend schema_sanitizer.py to handle both OpenAI and Responses tool shapes.
• Call strip_pattern_and_format() inside the xAI non-streaming Responses path before responses.create().
• Add regression tests for Responses-format tools containing pattern/format.

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR